Large language model
The technology behind ChatGPT, Claude, and similar tools reads and writes human language. What that means for your business: a 90-minute meeting transcript becomes a 30-second summary. A prospect profile becomes a personalized email draft. A plain-English question becomes a database query. But the model alone is a very articulate generalist that knows nothing about your operations. Connect it to your structured data with a semantic layer and it becomes a domain expert. The model is commodity — several vendors offer capable ones. What you connect it to is what creates value.
Go deeper
Picture this: your regional director calls and asks why the Southeast region missed its revenue target. Today, someone pulls a report, stares at pivot tables, calls two branch managers, and emails you a summary three days later. With an LLM connected to your operational data, you ask the question out loud and get: 'Southeast revenue was $180K under target. Root cause: two technicians resigned in August, reducing completed jobs by 23%. Remaining techs maintained their close rate, so this is a capacity issue, not a performance issue. Comparable regions backfilled in 14 days on average.' That answer took nine seconds.
The trap is treating the LLM itself as the product. Every vendor has access to roughly the same foundational models. The difference is what the model is connected to. An LLM with access to your governed data, your metric definitions, and your business rules is a different animal than the same LLM answering from general knowledge. When evaluating vendors, ignore the model name — ask what data layer sits between the model and your questions.
Questions to ask a vendor: Which LLM do you use, and what happens when you switch to a newer one — does our setup break? Can we see exactly which data sources the model queries when answering a question? What happens when the model does not have enough data to answer confidently — does it say so, or does it guess?